160 likes | 297 Views
Calorimeter Software Meeting 22 / 09 / 2004. Calo-related Activity in HLT. Olivier Deschamps, Cristina C â rloganu, LPC / Clermont-Ferrand. Goal. Provide a fast reconstruction software of the calorimeter data for HLT purpose → photons, neutral pions & electron-ID.
E N D
Calorimeter Software Meeting 22 / 09 / 2004 Calo-related Activity in HLT Olivier Deschamps, Cristina Cârloganu, LPC / Clermont-Ferrand
Goal Provide a fast reconstruction software of the calorimeter data for HLT purpose → photons, neutral pions & electron-ID speed > precision efficiency > purity { Privilegiate NB : all execution times mentionned in this talk are averaged over a sample of minibias events passing L0+L1 on a 1GHz bi-processor
Two possible approaches • 1- Re-use the existing offline architecture : CaloDigits→CaloClusters→CaloParticles→CaloHypos→ProtoParticles→DV Particles Currently requires ~250 ms (without RawBuffer ) • 2- Alternative approach adopted for this talk • Simplified architecture based on few short and easy-to-maintain algorithms reproducing at best the offline performance : RawBuffer → TrgCaloClusters→TrgCaloParticles
TrgCaloClusterFromRawBuffer TrgCaloClusterCorrections TrgTrClusterMatch TrgPhotonMaker TrgResolvedPi0Maker TrgCalo sequence
Clusterisation • Use an adapted version of the offline algorithm : • Ecal digits directly converted from RawBuffer • Creation and storage of CaloDigits suppressed • Replace CA by 3x3 clustering • Larger clusters also possible if needed for merged π0’s • New TrgCaloClusters • CaloPosition(E,x,y,z), seed CellID, Prs/Spd in front, • label to distinguish neutral / brems / charged • Barycenter computation included in the algorithm • does not compute covariance matrices (2ms in Brunel) • Simplified method used to share the energy of cells common to several clusters • … the iterative Brunel algorithm for that purpose needs 5 ms • … bug fix for clusters on region boundary • Code speed up by a factor 2 since last presentation <multiplicity> : 90 clusters per L0/L1 minibias event Mean execution time : 3.8 ms
EM Corrections • Ecal & Prs energy leakage, S-shape & Longitudinal corrections • The main one slightly simplified compared to offline • -> Degradation at % level • -> it takes ~1 ms longer if using the ofline corrections • New generic tools that can be used for both HLT & offline • Identical for all clusters (neutral or charged) as in DaVinci Mean execution time : 1.2 ms
Clusters-tracks matching • Labels Neutral Clusters and Electrons • extrapolate TrgTracks from State == AfterT → Calo • compute distance on Ecal face for all cluster-track pairs • ’Neutral cluster’ if min distance > d1→ Used for photon Id • ‘Charged cluster’ if min distance < d2 → Used for electron-Id • default cut d1 = d2 = 0.6 calo cell size • check if TrgTrack compatible with Electron <multiplicity>: 80 ‘neutrals’ + 10 ‘charged’ ( 3 ‘electrons’ ) Mean execution time : 0.8 ms • Make the subsequent γ/π°-makers alg. faster by 1ms when looking to neutrals only wrt to considering all clusters !
default cut ‘Charged’ ‘Neutrals’ Clusters-tracks matching Neutrality estimator 1-e-dmin Used as Conf-Level for photon Id Offline-choice for separation: χ²>4 → Neutral χ²<40 → Charged With dmin = 0.4 → 98% overlap with the offline neutral clusters
Electron Id Offline e-Id based on cluster-track energy-position χ² → 40ms (with the full set of offline tracks) Simplified PID procedure for HLT: • Geometrical matching between tracks and charged clusters • E/p match • Prs response for the cluster projection • Correct for Bremmstralhung • → track extrapolation from origin • matching a (neutral) cluster
Electron Id With default cuts: 92% efficiency and 53% purity for J/psi electrons
Photons maker • PhotonMaker algorithm : • Simply get the neutral TrgCaloClusters • Cluster position → photon momentum (assume (0,0,0) origin) • Optional cuts : Pt & ‘Pid-CL’ (default 200 MeV & dmin > 0.4 cells) • Produce TrgCaloParticles (Trg/Calo/TrgPhotons/) with : • 4-momentum, Id (22), ‘PId-CL’ (1-e-dmin), TrgCaloCluster CellID TrgCaloParticle class definition available in the next LHCB/Event package release <multiplicity> : 50 photons per L0/L1 minibias event Mean execution time : 0.3 ms Eoffline/Eonline = 1.0035, RMS=0.009
Resolved π°s maker • ResolvedPi0Maker algorithm : • Loop on all photon pairs • Addphotons4-momentum • Optional cuts : Pt (default : 200MeV/c) • Mass window (default : [100-160] MeV/c²) • ‘Pid-CL’ = (CLγ1* CLγ2), (default : no cut) • Produce TrgCaloParticles (Trg/Calo/TrgResolvedPi0/) with : • 4-momentum, Id (111), ‘Pid-CL’ (CLγ1* CLγ2), TrgCaloClusters pair CellID <multiplicity> : 30 π°s per L0/L1 minibias event Mean execution time : 0.8 ms
Resolved π°s performance Mass spectrum using Bd->3π sample Pt(π°) > 1 GeV/c clear π° peak ! Mass resolution ~ 10 MeV/c² similar to offline π°s !
Merged π°s • MergedPi0Maker algorithm : • Re-using offline iterative algorithm -> 32 ms (15 iterations) • In practice few iterations are enough → 11 ms (5 iterations) • Pt needs to be at least ~2 GeV for a π° to be merged in LHCb • Requiring Pt>1.5 GeV the algorithm essentially never runs → 0.5 ms However, merged pi0s maybe not usefull for HLT purpose. Could be considered inclusively with high Pt photons → simplify the clusterisation step (only 1 pass needed, no need to detail individual cell energy within TrgCaloClusters ) • Will be implemented only if required by physics (B→3π)
Overall speed performance Mean time for the whole TrgCalo sequence : ~7.5 ms Whole sequence mean Execution time (ms) Clusterisation Corrections ClusterTrMatch resolved π° Photons Number of clusters Execution time (ms)
Conclusions & perspectives • First version of fast CaloReco software for HLT exists • Few short and easy-to-maintain algorithms • Mean execution time ~ 15 ms within budget • Provide photon, resolved π° and electron-Id • HLT π°s have essentialy same resolution as offline • C++ optimisation / electron-Id alg improvement going on • To be done : • - Validate performance, efficiency/purity, overlap with offline reco. • - Develop physics HLT selection (B→K*γ , π+π-π°) to check performance.